Supply Chain

What Is Forecast Bias, and Why Is It Worse Than Forecast Error?

A forecast that is wrong in both directions is annoying. A forecast that is wrong in one direction is expensive, because every downstream buffer silently absorbs the same mistake over and over.

Forecast bias is the average signed error: the sum of forecast minus actual across periods, divided by the number of periods. Note the word signed. Forecast error metrics like MAPE take absolute values, which deliberately throws away direction, and direction is the whole story. A forecast that misses by fifty units high and fifty units low has zero bias and real error. A forecast that misses by forty units low six times running has substantial bias, and it is quietly draining inventory every single period.

The worked example

Take six periods with signed errors of -40, -35, -50, -30, -45, -38 units, all negative, meaning the forecast sat consistently below actual demand. Mean bias is -39.7 units per period. Mean absolute deviation is 39.7. The ratio of cumulative signed error to MAD is the tracking signal, here -6.0, and a common practitioner rule treats a tracking signal beyond roughly plus or minus four as evidence the forecast is systematically off rather than merely noisy.

Why bias compounds and error does not

Random error is absorbed by safety stock, which exists precisely for that purpose, and over time the misses offset. Bias is not absorbed, it accumulates. A consistently low forecast draws inventory down period after period until a stockout, and the stockout gets attributed to a supplier or a demand spike rather than to the forecast that caused it. A consistently high forecast does the reverse, building stock that eventually gets written off as obsolete, again with the blame landing somewhere else.

Safety stock is designed to absorb random error. It has no defence at all against a forecast that is wrong in the same direction every month.

Where bias comes from is usually organizational

Systematic bias is rarely a modelling failure. It is usually incentive. Sales forecasts skew optimistic where the forecast doubles as a target. Operations forecasts skew conservative where a stockout is punished harder than excess stock. Consensus forecasting processes often average these into something biased in whichever direction the more forceful function sits. The diagnostic is simple: measure bias separately for each contributing function, and the source usually identifies itself within a quarter.

What to do once you find it

Bias is the easiest forecast problem to fix, because a known systematic offset can simply be corrected. If the tracking signal shows a persistent negative bias of forty units, adding forty units is a legitimate and immediate improvement. Do it visibly rather than silently, though, because the more durable fix is removing whatever incentive produced the skew. A correction factor that nobody discusses tends to get forgotten and then double-counted when the underlying behaviour changes.

Use the forecast bias calculator to compute mean bias and tracking signal on your own history. Measure your bias

Published 2026-08-08.